Using AI In Audio Debates & GR Research [Video]

[Sal1950]

When it improves it’s not going to be better for us , it will be better at deception and following the agendas setup by thier owners we will never know truth again.

Agreed. But I'm an ole geezer that has maybe watched too much Battlestar Galactica. LOL
I'll be long gone before that I hope.

 

[Axo1989]

I think we all agree LLMs are not for mission critical tasks. I doubt anyone here is involved in mission critical tasks. …

Ok?

But I’d say when clients rely on our services to meet their statutory obligations and perhaps defend their operations against legal challenges, they may beg to differ.

… This is last I gonna post on this thread …

Ok.

 

[RandomEar]

The scary thing here is that multiple members are right or were right, depending on when they made their experience with LLMs. The ability of LLMs to successfully complete a task has roughly doubled every 7 months (!) in the past 5 years. In the past 2 years, that learning rate has possibly be even slightly higher.

So depending on when you last gave an LLM a challenging task from your field of expertise, your impression of how well it did might already be way out of date and you could be judging today's models far too pessimistic.

That being said, I agree with the general consensus that "AIs" will not make you smart: You should have a good general understanding of the topic you're asking it questions on to judge if the answer just sounds convincing or if it is actually correct. The models will always keep making egregious mistakes that are hard to spot for non-experts, but that error rate will also continue to decrease. Likely up to a certain, currently unknown point but never to zero - at least with the current transformer model topologies.

 

[Axo1989]

The scary thing here is that multiple members are right or were right, depending on when they made their experience with LLMs. The ability of LLMs to successfully complete a task has roughly doubled every 7 months (!) in the past 5 years. In the past 2 years, that learning rate has possibly be even slightly higher.

So depending on when you last gave an LLM a challenging task from your field of expertise, your impression of how well it did might already be way out of date and you could be judging today's models far too pessimistic. ...

Haha, totally looking forward to field identification of Australian grass species from sparse fragments of drought-desiccated inflorescence. Otoh the metric in that paper is 'length of tasks Als can do' roughly doubling every 7 months if I'm reading correctly (but I may have missed something).

 

[Alan O]

Danny is the perfect person to start off with and then jump away from.

I agree with the second half of that sentiment

I like to give everyone a fair shot, and when I first saw GR Research on YouTube a few years ago I started watching - it was only about crossover component 'upgrades' at the time, iirc. I was sceptical about super expensive capacitors, but I thought the idea of replacing electrolytics with film caps might have some merit.

I went ahead and did that with a modest pair of speakers I had - using some inexpensive film caps I happened to have -- and I heard a noticeable improvement (where I'm being deliberately careful to say I heard an improvement rather than there was one ). Annoyingly, it only struck me afterwards that it would have made sense to change one speaker first so I could then do direct A/B comparisons.

Anyway, even if my capacitor replacement did really make a difference, there was one factor I'd overlooked. I hadn't simply replaced electrolytic caps with film caps - I'd replaced 30-year-old caps with new caps. So there was no way to decide what might have made a difference.

Then some time later Danny started selling cables, with some silly demonstration of how a speaker cable used as an FM antenna can work as an FM antenna. And then his mains cables... And at that point I was 50/50 on him... many audio folk manage to mix sense with nonsense, so maybe there was still some sense worth listening for?

But the final straw was his recent speaker products, which employ some nonsense from Shunyata Research... and I'll have nothing to do with anybody who pushes BS like that. (And, also, I was rather disappointed that Ron - can't remember his channel - was in on it, as I'd had a pretty high opinion of him after some of the good stuff he'd previously done).

My experience with Paul McGowan at PS Audio was similar. When I first came across him he seemed like a helpful expert. But as he edged more towards subjective claims he couldn't back up with objective evidence, I went off him too. And I completely gave up on him when he pushed something (I can't remember what) that I knew was pure woo-woo. (I was just reminded of him when I watched Amir's YouTube take-down of some phase shift nonsense Paul had been going on about - YT popping up old ones can come in handy sometimes )

Anyway, that's my long-winded way of saying: Give everyone a listen, but only stick with people who can back up claims with actual science.

 

[338h10]

The scary thing here is that multiple members are right or were right, depending on when they made their experience with LLMs. The ability of LLMs to successfully complete a task has roughly doubled every 7 months (!) in the past 5 years. In the past 2 years, that learning rate has possibly be even slightly higher.

So depending on when you last gave an LLM a challenging task from your field of expertise, your impression of how well it did might already be way out of date and you could be judging today's models far too pessimistic.

That being said, I agree with the general consensus that "AIs" will not make you smart: You should have a good general understanding of the topic you're asking it questions on to judge if the answer just sounds convincing or if it is actually correct. The models will always keep making egregious mistakes that are hard to spot for non-experts, but that error rate will also continue to decrease. Likely up to a certain, currently unknown point but never to zero - at least with the current transformer model topologies.

The time when we know AI has dominated is when our judicial/arbitration system relies on AI for neutrality -- for example, when a transaction can have multiple reasonable tax treatments (depending on which side you are on and which law firm/Big4 you hire), and the judge says that everyone has a good point, and I don't want to side with one party, but let's see what the AI has to say because it is neutral ... This day could come in 10-20 years.

 

[Digital_Thor]

It is called Gemini from Google but it is running in fast mode, designed to answer quickly. You can change it to thinking mode which slows it down and provide better answers. I find Chatgpt paid version to be superior however.

I am sorry to disappoint, but this is just an AI image based on what I would like this room to look like. Just thinking ahead.

Haha... I figured that out as soon as I saw the chair pointing in the 'wrong' direction - then I knew it was just for show

 

[xnor]

Btw, anyone paying for these LLMs is also supporting the destruction of the PC market.
You'll own nothing and be at the mercy of a handful of global conglomerates not jacking up subscription prices, which they will.

I've also heard how people are very happy with AI data centers being built near them. (/s)

 

[DatSpeakerTho]

People who only use the free AI services for casual tasks have no clue what they are up against. For AI to be useful one has to know where they are going and be able to recognize when they get there. It also involves carefully prompting to make sure you don't take a contorted route that doesn't end up wasting a lot of time. A lot of times telling it what you don't want us just as important and that requires some knowledge.

So basically, you already need to understand a topic before using AI so you can spot errors which begs the question, why use it in the first place lol.

Every time I've used AI for anything I'm left wondering what anyone is getting out of it.

 

[xray]

So basically, you already need to understand a topic before using AI so you can spot errors which begs the question, why use it in the first place lol.

Every time I've used AI for anything I'm left wondering what anyone is getting out of it.

I use AI daily at work. It's good for gathering information, summarizing information, performing tasks and potentially interpreting information and making recommendations. AI can make mistakes and if you are trusting it to act on your behalf without being in a position to verify it's correct then it's a big risk.
Having said that, my experience has been that AI probably is saving me 3 to 6 hours a day, depending on the task at hand. It's good at examining my previous work and following that precedent for new work. I still need to correct it for certain details and nuances.

 

[testp]

So basically, you already need to understand a topic before using AI so you can spot errors which begs the question, why use it in the first place lol.

Every time I've used AI for anything I'm left wondering what anyone is getting out of it.

if we're talking about small nuances in subjective/objective audio sure, but if you would like to research about more fact based topics, AI is a good gatherer of information. just the other day i was looking for a psu trigger in europe, could not find any.. asked AI, got 3 options.
also little while ago, wanted to know the width of the back of the stand that is dapered (no specs for that, just pictures), when they arrived, sure enough it got it correct.

for an instrument AI is really good

 

[DatSpeakerTho]

if we're talking about small nuances in subjective/objective audio sure, but if you would like to research about more fact based topics, AI is a good gatherer of information. just the other day i was looking for a psu trigger in europe, could not find any.. asked AI, got 3 options. all that is really good

I asked it come up with a really basic pcb, gave it everything it should need but it failed miserably. Just spent its entire time telling me how great its going to be but it just needs one more bit of info. Kind of gets old after it does tha 10 times. I made the pcb schematic myself in a few minutes.

I just want to keep my critical thinking skills so I avoid it.

 

[DatSpeakerTho]

AI can make mistakes and if you are trusting it to act on your behalf without being in a position to verify it's correct then it's a big risk.

So like I said, you already need to know what youre doing to know when its wrong, which again begs the same question of why bother?

 

[testp]

I asked it come up with a really basic pcb, gave it everything it should need but it failed miserably. Just spent its entire time telling me how great its going to be but it just needs one more bit of info. Kind of gets old after it does tha 10 times. I made the pcb schematic myself in a few minutes.

I just want to keep my critical thinking skills so I avoid it.

if your talking about circuits, then that still isn't it's best feature (hidden paths aren't learned), if we would look at stackoverflow new topic frequencies regarding AI related vs. not AI related topics that are in down-trend, circuits aren't affected proven my econometric models (i made a school paper on internet traffic: before vs. after AI)

 

[xray]

So like I said, you already need to know what youre doing to know when its wrong, which again begs the same question of why bother?

I guess you missed the part where I said it saves me 3 to 6 hours a day of work?

 

[Axo1989]

I guess you missed the part where I said it saves me 3 to 6 hours a day of work?

Assuming you worked a 7 hour day (say) I’m glad I wasn’t paying you for those days you were wasting 6 hours on stuff a stochastic parrot could wing it through.

 

[xray]

Assuming you worked a 7 hour day (say) I’m glad I wasn’t paying you for those days you were wasting 6 hours on stuff a stochastic parrot could wing it through.

The lower tier free services are parrots. The advanced paid services are in a different league altogether. One valuable aspect of what AI can do is free you from time consuming grunt work like documenting changes, creating/updating user manuals for new functionality, managing deployments, etc. That's about 50% of the time savings for me. Of course what it produces still has to be vetted and tweaked by someone knowledgeable.

 

[DatSpeakerTho]

I guess you missed the part where I said it saves me 3 to 6 hours a day of work?

Nah I just don't care because thats not relevant to me at all. You could save a million working hours and id still think AI is shit.

I genuinely feel like a loser when I use AI. I like to learn things and actually understand how they work. Ive probably learned more from the process of correcting AI responses than I have from AI.

I find it funny that apparently one needs to pay money to get the useful AI? Like you have to actively financially contrcontribute to the destruction of so many things, just to get responses that arent trash? Well, at least it sorted some documents nicely for you while the data centers ruined the areas they were constructed it. i guess its pretty human to want to make our petty jobs a bit easier at a massive cost. Whateve, no one cares. Sort them files.

 

[Sal1950]

Having said that, my experience has been that AI probably is saving me 3 to 6 hours a day, depending on the task at hand.

Maybe just my experience but still, in pre AI times, I found a properly worded Google search to be mostly better at giving a correct answer than AI does today.
YMMV

 

[xray]

I find it funny that apparently one needs to pay money to get the useful AI?

Interesting perspective.
So you have the expectation that you are entitled to have something useful for free?